Hierarchical Disease-State Generators for Neurodegenerative Genomics: A Benchmark Proposal for Intervention-Conditioned Multi-omic Generation

David Scott Lewis, Enrique Zueco

Track: Full / long paper (5-8 pages)

Keywords: single-cell -omics, multi-omic integration, intervention-conditioned generation, multi-omic diffusion, perturbation effect modeling, conditional generation, cross-modal generation, disease-state generators, neurodegeneration, Alzheimer’s disease, Parkinson’s disease, counterfactual cell states, CRISPR perturbations, drug perturbations, regulatory edits, enhancer–TF–gene hierarchy, gene regulatory networks, mechanistic constraints, hierarchy fidelity metrics, cross-context generalization, uncertainty quantification, conformal prediction, latent diffusion models, multimodal latent encoders, perturbation effect prediction, counterfactual planning, benchmark design

TL;DR: Intervention-conditioned omic diffusion for AD/PD generates counterfactual cell states with enhancer→TF→gene priors and calibrated uncertainty; benchmark tests cover hierarchy fidelity, perturbation prediction, context transfer, and planning.

Abstract: Genomics is increasingly framed as targeted engineering of cellular and tissue states, yet current generative models often lack biologically grounded evaluation and mechanistic constraints. Neurodegeneration provides a concrete stress test: single-cell atlases reveal cell-type–specific vulnerability, microglial activation programs, reactive astrocyte states, and neuronal stress-response trajectories that should be generatable, intervenable, and verifiable. We propose the disease-state generator task: conditional generation of transcriptomic and epigenomic cell states under interventions (drugs, CRISPR perturbations, or regulatory edits), with uncertainty-aware outputs and hierarchy-aware acceptance criteria. We propose (i) a practical architecture—multimodal latent encoders coupled to conditional diffusion—augmented by (ii) hierarchical regulatory priors spanning enhancer→TF→gene structure, and (iii) an evaluation suite organized around workshop “barriers and frontiers”: regulatory-hierarchy fidelity, perturbation prediction, cross-context generalization, and uncertainty-calibrated intervention ranking. The framework doubles as an evaluation surface for DNA foundation models (Enformer, Nucleotide Transformer, DNABERT): hierarchy priors and sequence-derived regulatory features provide a task-aligned interface for probing whether FM-learned representations improve counterfactual generation. The draft is deliberately specification-gated: it defines formal acceptance criteria—uncertainty thresholds, hierarchy-consistency bounds, and coverage guarantees—that determine when generated counterfactuals are trustworthy enough for downstream use, making “state engineering” measurable in AD/PD-relevant settings.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Submission Number: 78